Human-architected, AI-implemented: I design, agents code. What that means
Fleet-wide questions, AI trajectory takes, roadmap ideas? Discussions — not tied to one repo's tools, that's what this one is for.
Retired software engineer, 9th District Vienna. I wrap everyday tools in a five-layer fleet pattern, then snap repos together into pipelines. 190+ small repos, one cohesive architecture.
Start here: Why so many repos? — plain-language value prop, 3 minutes.
flowchart LR
W[Wrapee<br/>tool or device]
M[MCP server<br/>agents]
B[Webapp<br/>human dashboard]
D[Winapp<br/>Tauri NSIS .exe]
A[AI chat and workflows<br/>skills plus agentic jobs]
W --> M
M --> B
M --> D
B --> A
Each repo ships four layers wired to one wrapee: an MCP server for Claude or Cursor, webapp and winapp for humans, and AI specialist chat and workflows in the web stack. The wrapee itself isn't shipped — you already have it, or the repo helps you find, install, and wire it in (from Inkscape and Lean to robot cars and audio gear).
| Sphere of Interest | Focus & Architecture | Detail Sub-Readmes |
|---|---|---|
| Meshed MCP Fleet | 190+ repos wrapping tools across robotics, CAD/3D, audio/video, Calibre, and transit into a 5-layer standard. Repos run completely standalone or snap together into 1-hop crossconnect pipelines with soft degradation. | Why So Many Repos? Full MCP Catalog Fleet Crossconnects |
Rustification (r11n) |
The next fleet regime: migrating from Python/FastMCP to pure-Rust workspaces (<domain>-core, <domain>-mcp, <domain>-app). Single cargo toolchain, sub-millisecond execution, zero node_modules, and zero Python runtime drift. |
Rustification SOP Dev Stack |
| ArtCraft 3-Tier Scaffolding | 2026-10-17 Red-Letter standard replacing multi-process IPC: L1 core (pure algorithmic logic) → L2 CLI + MCP (headless agent tool surface) → L3 in-process Tauri (zero-port GUI). Clean modular boundaries with mechanical verification gates. |
Dev Stack Architecture Standards Fleet |
| Stigmergic Swarm & r11n | Autonomous multi-agent coordination without central orchestrator bottlenecks. Agents coordinate stigmergically via shared environment traces (.r11n/claim.json) and declarative invariant reconciliation (Observe -> Diff -> Act). Ephemeral workers run in space-efficient WSL2 micro-sandboxes (OverlayFS + Git Worktrees + dynamic memory reclaim) rather than heavy, RAM-eating standalone VMs. |
Stigmergic Swarm & WSL2 Sandboxing Fritz Universal Agent |
| Local LLM Sovereign Inference | 100% local reasoning on Goliath (RTX 4090) with Bonsai 2 27B PQ2_0 and Glimmer 27B Q4. Prompt truncation proxy, aggressive KV caching, and strict $0/EUR token targets. Cloud models are an emergency fallback only (<$20/mo budget) for 1M-context sweeps. | Local LLM Stack Coming Next & Hardware Workshop (Goliath) |
| Own "Craft Apps" (GameCraft, DemoCraft) | Autonomous generative engines: describe a game in a single sentence and get a validated, headless-tested Godot 4 project (gamecraft); scene & procedural generation (democraft-mcp); visual design exploration (lookcraft). Creative sandboxes built for agents and humans. |
GameCraft Project Make / Move / Meet |
| AI-Reversing & Hardware Revival | Static binary analysis, Ghidra decompilation over MCP (via ReVa), Directmedia format reversing, and automated unpacking to revive legacy hardware and retro systems. Currently active via reversing-mcp (revops), transitioning to pure-Rust reversecraft. |
Reversing MCP Sneaky Feats |
- One wrapee per repo, five layers. Each repo wraps one app or device you already use (Blender, QCAD, Plex, a Yahboom robot car, Vienna transit): MCP tools for agents, a dark web dashboard, a Tauri Windows app (NSIS .exe), plus in-app AI chat fed by domain skills and agentic workflows.
- Human faces are not decoration. Agents act, humans approve. Watch telemetry, sign off on posts and destructive ops, hit the big red stop button.
- Every repo has intrinsic AI. A chat page that ingests bespoke domain skills (ask yahboom-mcp about robot types in plain words) plus agentic workflows that run multi-step jobs locally. Details: Why the fleet looks like this.
- They snap together, optionally. Any repo runs alone. When two repos know each other they offer a 1-hop link (CAD file to 3D print to robot mount to VR twin to spoken "done"). No forced installs, friendly error if the peer is missing.
Details: Why the fleet looks like this — why one repo per tool, why a webapp each, what crossconnects do, and an honest note on stars.
| If you want... | Go to |
|---|---|
| Design things, move robots, meet in VR (the interlock chain) | Make / Move / Meet — Blender, QCAD, FreeCAD, Resonite VR, Yahboom, Dreame, Nori, Unitree |
| Sound, video, DJ sets, knowledge, everyday Vienna life | Play / Know / Home — Mixxx, stems, Resolume, memory, Calibre, arXiv, Plex, transit |
| Every server in one registry | Full MCP Catalog — 126+ repos by category |
| How repos link (clickable graph + rules) | Fleet Crossconnects — pipelines, 1-hop contract, soft degradation |
| Two deep dives, kept separate on purpose | Fleet Highlights (mobile-DJ video fork, Lean proof loop) — Coming Next (local sovereign agent, 0 EUR per token) |
| The machine and the hardware behind it | Workshop — Goliath, RTX 4090, Boomy, Bumi |
| Who runs this (workshop photo, Benny, follows, thanks) | About Sandra — the personal tail, kept off this page on purpose |
| How it is built (developers: tools, tests, standards, CI) | Dev Stack — very technical, everyone else stop here |
| Why the fleet is being rewritten in Rust (r11n) | r11n — the why, the SOP, the first six converts |
| Stigmergic swarm & WSL2 micro-sandboxes | Stigmergic Swarm — autonomous repo repair without VM bloat |
| The rogues' gallery (Alexa air-gap, listening ops, live demos) | Sneaky Feats — mischief with guardrails |
| So you want to be a v̶i̶b̶e̶c̶o̶d̶e̶r̶ "agentic architect" | Vibecoding Notes — subscription round-robin, DeepSeek peak-hour gotchas |
| The wins that actually cleared a real bar | Our (Nontrivial) Successes — external validation, hard problems solved, real thresholds crossed |